Volcanic engines release large-scale bean pack models in series 2.1 consisting of flagship Doubao-Seed-21-Pro and light Doubao-Seed-21-Turbo, on board the volcano ark. At the same time, the video generation model Seedance 2.5 is expected to be officially released in early July, the audio generation model 1.0 has also been opened and the product line has further expanded from the language model to video and audio generation.

Call and market share Disclosure

The president of the volcano engine Tan at the launch revealed that, as of June this year, the daily average of the Bean Pack model Token calls had exceeded 180 trillion, 10 times more than last year. At the same time, the company claims that the market share of the volcanic engine is 49.5 per cent in the Maas service market in China.

This set of data reflects the continued rapid growth in domestic business demand for large model deployments, as well as the fact that the volcanic engine is focusing on production-level scenarios, not just demonstration of model capabilities.

Pricing for enterprise-level deployment

The large bean bag model 2.1 Pro is priced at $6 and $30 per million Token input; the input price can be reduced to $1.2 under CL. Volcanic engines claim that in Coding and Agent scenarios, combined costs can be reduced to $ 1.96 per million Token.

The Turbo version, on the other hand, further downwards the price, close to the Pro capability, and is better suited to the HF call scene. Volcanic engines have also launched a continuous iterative version of Doubao-Seed-Evolving, and businesses can be updated with follow-up models without replacing API access nodes.

Focus on CDing and Agent

Volcanic engines focused this time on programming and Agent capabilities. According to the company, the Bean Packed Large Model 2.1 Pro demonstrated close to or exceeding international models such as OpenAI GPT-5.5, Anthropic Claude Opus 4.7 in several programming and mission performance baseline tests.

In the field presentation, the volcano engine gave a case of chip design RTL. According to the company, the model was running for nearly 18 hours on a continuous basis, after nine rotations, completing the production of six core modules and 1303 line RTL codes and using a simulation, testing and integrated inspection process.

In the Agent scene, the volcanic engine also showed multi-agent collaboration cases. According to its version, a developer deployed more than 500 Agents, triggering thousands of calls of tools and eventually completing the construction of a large 3D urban map.

Multimodel products continue to expand.

In addition to the language model, the Volcanic Engine predicts that Seedance 2.5 will be officially launched in early July. The video model supports up to 30 seconds of single-part video generation and up to 50 joint input of all-modular materials. According to the company, the new version had been further enhanced in terms of camera consistency and video editing control.

The bean-cover audio generation model 1.0, published on the same day, supports text and reference audio input and directly generates complete audio content that contains multi-role white, background music and environmental sound effects. The model supports 2 minutes of audio creation one time, and the API has already opened an invitation in the volcano ark.

Accessed office and development tool

Volcanic engines indicate that the Bean Pack megamodel 2.1 has access to partners such as WPS, Access, Unity, and compatible with the development frameworks such as Claude Code, Codex and tools such as TRAE and buttons.

In terms of product tempo, price and range of access, the volcanic engine is moving to a larger business acquisition and productivity scenario. As video and audio models continue to be online, domestic large model competition is shifting from a single text capability to a multimodular generation and actual delivery capability.